Here is the foundational myth of modern science: when a finding is wrong, the system catches it. Peer review flags errors. Replication weeds out flukes. Citations reward truth and punish falsehood. Self-correction is not a feature of science — it is science.
The evidence says the opposite. The system that propagates scientific knowledge — the citation network — doesn't just fail to filter out bad science. It preferentially amplifies it.
The 153x Problem
In 2021, Marta Serra-Garcia and Uri Gneezy published a study in Science Advances that should have detonated across every scientific discipline. They examined papers from psychology's replication crisis — studies that had been formally tested and found not to replicate — and tracked their citation patterns.
Non-replicable papers were cited 153 times more than replicable ones.
In Nature and Science specifically? 300 times more.
This isn't a rounding error. This is the citation system functioning as a falsehood amplifier. The findings most likely to be wrong are the findings most likely to spread.
And it gets worse. When Serra-Garcia and Gneezy looked at what happened after studies were shown to be non-replicable, they found that only 12% of subsequent citations even acknowledged the replication failure. The other 88% cited the discredited work as if nothing had happened.
The Undead Literature
Retraction is supposed to be the nuclear option — the final correction when a paper is so flawed it needs to be removed from the record. It should stop the spread. It doesn't.
A systematic analysis of 13,252 citation contexts across a 60-year database found that only 5.4% of post-retraction citations acknowledged the retraction. The rest cited retracted papers as valid science.
The case studies are staggering. The social psychologist Diederik Stapel fabricated data across dozens of papers. In January 2026, Science reported that of 88 papers citing Stapel's retracted work, 39 — nearly half — had drawn conclusions that were substantially weakened without the fraudulent foundation. Journals had flagged just four of them.
A 2025 study of another retracted social psychologist's work found that 58.28% of all citations came after retraction. Some papers received more citations dead than alive.
Researchers have started calling these "zombie papers." A March 2026 study in Research Policy analyzed 25,480 retracted articles published between 1923 and 2023. Articles retracted for serious misconduct — fabrication, data manipulation — experienced the longest delays before retraction, giving them more time to embed in the literature. And the more citations an article accumulated before retraction, the longer it took to kill.
Even Wikipedia can't escape. An arXiv study from January 2026 found that 71.6% of Wikipedia citations to retracted papers were problematic — continuing to treat them as valid. Higher academic citation count predicted slower correction on Wikipedia. The academic prestige that got the paper cited in the first place now shields it from correction.
The Scale of the Infection
This would be concerning if we were talking about a few hundred bad papers. We're not.
The Retraction Watch database has grown to over 63,000 retractions as of March 2026. In 2023 alone, there were 13,000 — many from a single publisher, Hindawi, whose journals had been compromised at scale.
But retractions are just the visible tip. A March 2026 Northwestern University study by Richardson and Amaral found that fraudulent publications are now appearing faster than legitimate ones. Paper mill output is doubling every 1.5 years. Legitimate scientific output doubles every 15 years. Do the math on where those curves cross.
An AI scan of 2.6 million publications in cancer research flagged 9.87% as probable paper mill products. That rate grew from 0.06% in 1999 to 14.4% in 2024. In one medical subfield, nearly one in seven papers may be fabricated.
These aren't rogue scientists. Science investigations have documented organized criminal networks — paper mills that mass-produce manuscripts with fabricated data, brokers who sell authorship slots at up to $5,000 each, and agents planted on editorial boards to ensure acceptance. One Russian operation, 123mi.ru, claims to have brokered 20,000+ authorship slots across 4,000+ papers.
Conservative estimates suggest 400,000 fraudulent articles have infiltrated the scientific literature over the past two decades. Only about 7,275 retractions have been officially linked to paper mills. The detection rate is under 2%.
Registered Reports: The Experiment That Proved the Disease
There is one piece of evidence that makes the case definitive. It comes from an attempted fix.
Registered reports are a publishing format where journals review and accept a study's methodology before the results are known. The hypothesis, methods, and analysis plan are locked in advance. You can't massage data to reach significance. You can't bury null results.
When researchers adopted registered reports, the rate of statistically significant findings dropped from over 90% to roughly 50%.
Let that sink in. Under normal publishing conditions, more than 90% of studies find what they're looking for. When you remove the ability to adjust the analysis after seeing the data, the rate drops to what you'd expect from honest inquiry — some hypotheses confirmed, some not.
The gap between 90% and 50% is the inflation rate. It's the measure of how much the system distorts science on the way through.
But here's the meta-contradiction. Registered reports work — and almost nobody uses them. Of 36 experimental psychology journals that adopted the format, 11 published zero registered reports by the end of 2023. A single journal accounts for 30% of all registered reports ever published. Researchers avoid them because null findings damage careers. The incentive structure doesn't just tolerate the disease. It rewards it.
The Political Weaponization
Every genuine problem generates two kinds of exploitation. People who deny the problem exists, and people who use the problem to justify their own agenda.
The replication crisis is real. The evidence is overwhelming. And it is now being weaponized.
In the United States, the MAHA Commission and NIH director Jay Bhattacharya have cited replication failures to justify cutting NIH funding by 40%. The logic: if science can't replicate its own findings, why fund it?
This is a meta-contradiction. The fix for "not enough replication" is... cutting the funding that would pay for replication? As Brian Nosek, executive director of the Center for Open Science — the person who did more than anyone to expose the replication crisis — put it: "It's hard to read it as an investment in wanting science to improve."
The people most vocal about broken science are not proposing to fix it. They're proposing to defund it. The crisis is the pretext, not the concern.
What Self-Correction Actually Looks Like
Real fixes exist. They're just underfunded and undersupported.
Clarivate reformed the Journal Impact Factor in 2025 to exclude citations to retracted papers — the first systemic response from the citation infrastructure itself. The United2Act coalition — 40 participants from 15 countries — is coordinating detection and post-publication corrections. The nonprofit ERROR pays bounties for identifying serious flaws in published papers: more severe error, larger payout. Crowdsourcing integrity because institutions won't fund it.
These are good starts. They're also Band-Aids on a structural problem. The fundamental incentive hasn't changed: novel, surprising, statistically significant findings get published, cited, and funded. Replications, null results, and corrections do not. Until that changes, the amplifier stays on.
The Contradiction
Science claims self-correction as its defining virtue — the thing that separates it from ideology, superstition, and opinion. And at the level of principle, this is true. Science can self-correct. The tools exist. The methodology works when applied honestly.
But the actual mechanism through which scientific knowledge propagates — the citation system, the publishing incentives, the career structures — does the opposite. It preferentially amplifies unreliable findings. It ignores corrections. It rewards the disease and punishes the cure. And now, organized fraud has industrialized the exploitation of these structural weaknesses at a pace that outstrips the growth of legitimate science itself.
The most important claim science makes about itself — that it self-corrects — is contradicted by the system through which science actually spreads.
The knowledge is good. The plumbing is broken. And the plumbing is where everyone drinks.